Capital

Google Hedges Its AI Chip Bet With $12 Billion

Google handed Marvell a $12.2 billion stock warrant tied to future AI chip orders, and Broadcom's stock fell 5 percent the same day. Here is what the vesting terms actually promise.

CapitalBy Servola Tech Desk2026-08-194 min read

AI-assisted, edited by humans. | Editor-in-Chief: Leon Soliman | Editorial standards

Google Hedges Its AI Chip Bet With $12 Billion

Key takeaways

Google received a warrant for up to 58.97 million Marvell shares (roughly $12.2 billion, about 11.3 billion euros or 9.6 billion pounds, at the $206.58 strike price) on August 19, 2026 - but the shares vest only as Marvell books real revenue from Google's orders, not as a lump sum.
About 1.36 million shares vest automatically over the first year; the remaining 57.6 million vest in tranches of roughly 240,000 shares for every $500 million of qualifying Marvell revenue from Google, running through early 2033 - full vesting implies $120 billion in cumulative orders, a ceiling, not a spending pledge.
Marvell is not replacing Broadcom, which signed its own long-term TPU development agreement with Google in April 2026 through 2031; Marvell instead builds the AI inference accelerators, networking chips, storage and memory controllers that sit around Google's TPUs.
Marvell shares rose more than 6 percent and Broadcom fell more than 5 percent the same day - a market read that a second embedded supplier inside Google's custom-silicon supply chain is a real threat to Broadcom's share of future orders, not a symbolic gesture.
For any business running workloads on Google Cloud's TPU infrastructure, including in the EU and UK, a dual-sourced chip supply chain is a direct hedge against the capacity shortages and price spikes that have hit single-vendor AI chip buyers for the past two years.

A Warrant That Pays Out Only If Marvell Delivers

On August 19, 2026, Marvell Technology issued Alphabet's Google a warrant to buy up to 58.97 million Marvell shares at $206.58 each, exercisable until August 18, 2033. At full exercise, that is roughly $12.2 billion - about 11.3 billion euros, or 9.6 billion pounds - Google would pay Marvell if it chose to buy every share. That distinction matters: the $12.2 billion is what Google would spend to exercise the option, not money Marvell is paying out or a chip order Google has committed to.

The structure is what makes this different from an ordinary strategic investment. Roughly 1.36 million shares vest automatically in equal quarterly installments over the first year. The remaining 57.6 million vest only as Marvell actually books qualifying revenue from Google's purchases, in tranches of about 240,000 shares for every $500 million billed, running from Marvell's third fiscal quarter of 2027 through the end of fiscal 2033. Full vesting would require $120 billion in cumulative Google orders over that period - a ceiling on the incentive, not a forecast of what will happen.

Two Suppliers, Two Different Jobs Inside the Same Chip Stack

Marvell is not taking over Google's Tensor Processing Unit design. Broadcom still holds that role: the two companies signed their own long-term agreement in April 2026 covering future TPU generations and AI-rack components through 2031, and Broadcom remains, in its own words, deeply entrenched in Google's AI silicon roadmap. What Marvell picks up instead is the silicon built around the TPU - AI inference accelerators, networking chips, storage controllers, memory interface controllers and near-memory computing technology that let a TPU cluster actually move and process data at scale.

That division of labor is the point. Google is not swapping vendors; it is adding a second, financially incentivized one inside the same ecosystem, for the components that sit next to - not inside - its flagship AI chip. It is a narrower slice of the stack, but a slice large enough that Marvell's own investors treated it as a genuine expansion of the company's cloud-AI business.

Why Broadcom's Stock Fell on Someone Else's Deal

The market did not read this as a footnote. Marvell shares rose more than 6.6 percent the day the deal was disclosed, closing near $230, while Broadcom fell more than 5.5 percent to roughly $359. Alphabet's own stock barely moved. Analysts covering the announcement were explicit about why: Marvell is not displacing Broadcom's core TPU role, but it is now positioned to capture a growing share of Google's future AI-silicon spending in the categories around that chip - spending that would otherwise have gone entirely to Broadcom or been built in-house.

Marvell reports its fiscal second-quarter earnings on August 27, 2026, which will be the first chance for investors to press management on how much near-term revenue the Google relationship is actually expected to generate, beyond the long-dated vesting math in the warrant itself.

An Infrastructure Race Behind a Model Race Google Is Not Winning

The timing lines up with a wider pattern in how Google is competing. Prediction markets tracking which company will field the best AI model by the end of 2026 give Google only about a 7 percent chance, against roughly 66 percent for Anthropic, 14.5 percent for xAI and 8 percent for OpenAI. Google is not the favorite on model quality right now, by that measure, and it knows it.

What Google can still control is the cost, speed and specialization of the compute those models run on. Locking in a second, incentive-aligned silicon supplier - one that only gets paid in equity as it proves it can deliver - is a bet on infrastructure economics doing work that model benchmarks currently are not. It is a hedge against losing the AI race on two fronts instead of one.

What Dual-Sourced Silicon Means for Buyers Outside the US

None of this is US-only. Any organization running production AI workloads on Google Cloud's TPU infrastructure - a growing number of EU and UK enterprises using Vertex AI or dedicated TPU pods for training and inference - is exposed to how much chip capacity Google can secure, at what price, and from how many independent sources. A hyperscaler dependent on a single silicon vendor is a hyperscaler whose customers inherit that vendor's capacity constraints and pricing power during a shortage.

A second embedded supplier does not eliminate that risk, but it meaningfully reduces it, and it mirrors the logic EU policymakers have been pushing on cloud and chip vendors for their own sovereignty reasons: concentration in one supplier is a structural weakness, not just a procurement inconvenience. Watching whether Marvell's Google-linked revenue actually shows up in its upcoming earnings calls is now a genuine capacity-planning signal for anyone budgeting AI compute costs into 2027 and beyond.

Frequently asked questions

Does Google have to pay Marvell $12.2 billion?

No. The $12.2 billion is the maximum Google would pay Marvell if it exercised every share in the warrant at the $206.58 strike price. It is the cost of buying the shares, not a chip order or a cash payment from Google to Marvell, and Google is not committing to spend that sum.

Is Marvell replacing Broadcom as Google's chip partner?

No. Broadcom signed its own long-term agreement with Google in April 2026 to develop future TPU generations and AI rack components through 2031. Marvell's role covers chips built around the TPU - inference accelerators, networking, storage and memory controllers - not the TPU itself.

What has to happen for Google to receive all 58.97 million shares?

About 1.36 million shares vest automatically over the first year. The remaining 57.6 million vest only as Marvell books qualifying revenue from Google purchases, in tranches of roughly 240,000 shares per $500 million, through early 2033. Full vesting implies $120 billion in cumulative orders - a ceiling on the structure, not a forecast.

Why should a company outside the US care about this deal?

Any organization running AI workloads on Google Cloud's TPU infrastructure is exposed to how much chip capacity Google can secure and at what cost. A second, incentive-aligned silicon supplier reduces Google's dependency on a single vendor's production capacity and pricing power, which is a direct hedge for customers against future shortages.

The headline number is $12.2 billion, but the structure is the actual story: Google did not write Marvell a check, it wrote an incentive contract that only pays out as Marvell proves it can win business Google would otherwise send entirely to Broadcom or build itself. That is what a hyperscaler does when it wants a real second option, not a backup plan filed in a drawer - and it is worth watching whether Nvidia's own hyperscaler customers start writing the same kind of contract with Nvidia's rivals.

MarvellGoogleAlphabetBroadcomTPUcustom siliconAI chips

The Nvidia Tax Just Found a Second Source

Meta's MTIA chip Iris enters production in September after clearing testing in six weeks. Why a hyperscaler's second source caps the price you pay for compute.

3 min read

AI Policy

Google Buys a Bankrupt Airline's Data for AI Training

Google paid $10 million for Spirit Airlines' internal emails, chats and payroll records in a bankruptcy sale - a precedent EU and UK businesses should read closely.

3 min read

Decisions

The Third Pivot Was the Warning Magic Leap Missed

Magic Leap filed a WARN notice for 193 layoffs effective 1 October 2026, its fourth strategic pivot after $4 billion raised. The lesson is the pattern, not the pivot.

4 min read

← All articles